fix deepspeed4science links (#4358)

docs/_pages/deepspeed4science.md

@@ -22,14 +22,14 @@ We are proud to introduce [new Megatron-DeepSpeed](https://github.com/microsoft/
 
 To see how the new Megatron-DeepSpeed helps enabling new system capabilities, such as training models with massive sequences length, please read our [tutorial](https://github.com/microsoft/Megatron-DeepSpeed/tree/main/examples_deepspeed/deepspeed4science/megatron_long_seq_support).
 
-Meanwhile, our new Megatron-DeepSpeed has been applied to genome-scale foundation model [GenSLMs](https://github.com/ramanathanlab/genslm), which is a 2022 [ACM Gordon Bell award](https://www.acm.org/media-center/2022/november/gordon-bell-special-prize-covid-research-2022) winning genome-scale language model from Argonne National Lab. To achieve their scientific goal, GenSLMs and similar models require very long sequence support for both training and inference that is beyond generic LLM's long-sequence strategies. By leveraging DeepSpeed4Science's new Megatron-DeepSpeed, GenSLMs team is able to train their 25B model with 512K sequence length, much longer than their original 42K sequence length. Detailed information about this application can be found at [our website](https://deepspeed4science.ai/). GenSLMs team also hosts an [example](https://github.com/ramanathanlab/genslm/tree/main/examples/long-sequences) about how to use DeepSpeed4Science in the GenSLMs repo.
+Meanwhile, our new Megatron-DeepSpeed has been applied to genome-scale foundation model [GenSLMs](https://github.com/ramanathanlab/genslm), which is a 2022 [ACM Gordon Bell award](https://www.acm.org/media-center/2022/november/gordon-bell-special-prize-covid-research-2022) winning genome-scale language model from Argonne National Lab. To achieve their scientific goal, GenSLMs and similar models require very long sequence support for both training and inference that is beyond generic LLM's long-sequence strategies. By leveraging DeepSpeed4Science's new Megatron-DeepSpeed, GenSLMs team is able to train their 25B model with 512K sequence length, much longer than their original 42K sequence length. Detailed information about the methodology can be found at [our website](https://deepspeed4science.ai/2023/09/18/model-showcase-genslms/). GenSLMs team also hosts an [example](https://github.com/ramanathanlab/genslm/tree/main/examples/long-sequences) about how to use DeepSpeed4Science in the GenSLMs repo.
 
 
 ## Memory-Efficient EvoformerAttention Kernels
 
 [Evoformer](https://www.nature.com/articles/s41586-021-03819-2) is a key building block for scientific models such as DeepMind's AlphaFold. However, EvoFormer's multiple sequence alignment (MSA) attention frequently runs into memory explosion problems during training/inference, such as in protein structure prediction models. Existing techniques such as FlashAttention cannot effectively support Evoformer because EvoFormerAttention uses row-wise/column-wise/triangle attention, which are different from standard Transformer self-attention and cross-attention that require custom optimizations. To mitigate the memory explosion problem, we introduce `DS4Sci_EvoformerAttention` kernels, a collection of kernels that improve the memory efficiency of variants of EvoFormer. `DS4Sci_EvoformerAttention` is easy-to-use. To see how you can use it, please refer to our [tutorial](/tutorials/ds4sci_evoformerattention/).
 
-`DS4Sci_EvoformerAttention` has already been applied to [OpenFold](https://github.com/aqlaboratory/openfold), which is a community reproduction of DeepMind's AlphaFold2 that makes it possible to train or finetune AlphaFold2 on new datasets. With DS4Sci_EvoformerAttention kernels, OpenFold team is able to reduce the peak memory requirement by 13x without accuracy loss. Detailed information about this application can be found at [our website](https://deepspeed4science.ai/).
+`DS4Sci_EvoformerAttention` has already been applied to [OpenFold](https://github.com/aqlaboratory/openfold), which is a community reproduction of DeepMind's AlphaFold2 that makes it possible to train or finetune AlphaFold2 on new datasets. With DS4Sci_EvoformerAttention kernels, OpenFold team is able to reduce the peak memory requirement by 13x without accuracy loss. Detailed information about the methodology can be found at [our website](https://deepspeed4science.ai/2023/09/18/model-showcase-openfold/).
 
 <!-- OpenFold team also hosts an [example](https://github.com/aqlaboratory/openfold/blob/main/tests/test_deepspeed_evo_attention.py) about how to use DS4Sci_EvoformerAttention in the OpenFold repo. -->
 

docs/_tutorials/ds4sci_evoformerattention.md

@@ -69,6 +69,6 @@ out = DS4Sci_EvoformerAttention(Q, K, V, [res_mask, right_edges])
 
 ### 4.1 DS4Sci_EvoformerAttention eliminates memory explosion problems for scaling Evoformer-centric structural biology models in OpenFold
 
-[OpenFold](https://github.com/aqlaboratory/openfold) is a community reproduction of DeepMind's AlphaFold2 that makes it possible to train or finetune AlphaFold2 on new datasets. Training AlphaFold2 incurs a memory explosion problem because it contains several custom Evoformer attention variants that manifest unusually large activations. By leveraging DeepSpeed4Science's DS4Sci_EvoformerAttention kernels, OpenFold team is able to reduce the peak memory requirement by 13x without accuracy loss. Detailed information about this application can be found at [our website](https://deepspeed4science.ai/).
+[OpenFold](https://github.com/aqlaboratory/openfold) is a community reproduction of DeepMind's AlphaFold2 that makes it possible to train or finetune AlphaFold2 on new datasets. Training AlphaFold2 incurs a memory explosion problem because it contains several custom Evoformer attention variants that manifest unusually large activations. By leveraging DeepSpeed4Science's DS4Sci_EvoformerAttention kernels, OpenFold team is able to reduce the peak memory requirement by 13x without accuracy loss. Detailed information about the methodology can be found at [our website](https://deepspeed4science.ai/2023/09/18/model-showcase-openfold/).
 
 <!-- OpenFold team also hosts an [example](https://github.com/aqlaboratory/openfold/blob/main/tests/test_deepspeed_evo_attention.py) about how to use DS4Sci_EvoformerAttention in the OpenFold repo. -->